Task-aware Retrieval with Instructions
Official repository: github.com/facebookresearch/tart
Model descriptions
facebook/tart-full-flan-t5-xl is a multi-task cross-encoder model trained via instruction-tuning on approximately 40 retrieval tasks, which is initialized with google/flan-t5-xl.
TART-full is a 1.5 billion cross-necoder and it can rerank top documents given a query and natural language instruction (e.g., find a Wikipedia paragraph that answers this question.). Experimental results on widely-used BEIR, LOTTE, and our new evaluation, X^2-Retrieval show that TART-full outperforms previous state-of-the-art methods by levaraging natural language instructions.
More details about modeling and training are in our paper: Task-aware Retrieval with Instructions.
Installation
git clone https://github.com/facebookresearch/tart
pip install -r requirements.txt
cd tart/TART
How to use?
TART-full can be loaded through our customized EncT5 model.
from src.modeling_enc_t5 import EncT5ForSequenceClassification
from src.tokenization_enc_t5 import EncT5Tokenizer
import torch
import torch.nn.functional as F
import numpy as np
# load TART full and tokenizer
model = EncT5ForSequenceClassification.from_pretrained("facebook/tart-full-flan-t5-xl")
tokenizer = EncT5Tokenizer.from_pretrained("facebook/tart-full-flan-t5-xl")
model.eval()
q = "What is the population of Tokyo?"
in_answer = "retrieve a passage that answers this question from Wikipedia"
p_1 = "The population of Japan's capital, Tokyo, dropped by about 48,600 people to just under 14 million at the start of 2022, the first decline since 1996, the metropolitan government reported Monday."
p_2 = "Tokyo, officially the Tokyo Metropolis (東京都, Tōkyō-to), is the capital and largest city of Japan."
# 1. TART-full can identify more relevant paragraph.
features = tokenizer(['{0} [SEP] {1}'.format(in_answer, q), '{0} [SEP] {1}'.format(in_answer, q)], [p_1, p_2], padding=True, truncation=True, return_tensors="pt")
with torch.no_grad():
scores = model(**features).logits
normalized_scores = [float(score[1]) for score in F.softmax(scores, dim=1)]
print([p_1, p_2][np.argmax(normalized_scores)]) # "The population of Japan's capital, Tokyo, dropped by about 48,600 people to just under 14 million ... "
# 2. TART-full can identify the document that is more relevant AND follows instructions.
in_sim = "You need to find duplicated questions in Wiki forum. Could you find a question that is similar to this question"
q_1 = "How many people live in Tokyo?"
features = tokenizer(['{0} [SEP] {1}'.format(in_sim, q), '{0} [SEP] {1}'.format(in_sim, q)], [p_1, q_1], padding=True, truncation=True, return_tensors="pt")
with torch.no_grad():
scores = model(**features).logits
normalized_scores = [float(score[1]) for score in F.softmax(scores, dim=1)]
print([p_1, q_1][np.argmax(normalized_scores)]) # "How many people live in Tokyo?"